Alice authentic held movement sample and reserved-world repair, 2026-11-03

Original New Game / Easy completed its natural opening and reached visible cave gameplay. The retained scheduler adapter held the actual completed movement return, and read-only captures established the engine-selected Player/state chain. The first decoder failure was concrete: native world contact 1022 is outside live record count 208, with capacity 1024. The earlier ordinary-entity-only contact check rejected it. No movement or jump followed. usableForFooting=false: two neutral samples, allocation lifetime continuity, the unsupported/airborne comparison and wind/plank/knife progress remain open.

Exact runtime source was f137c490fbf730795b6eb41d0f891932d030af27, including published fault retention 0c2bc2da5641695157703a778a377e6f44b16d74. All 5904 staged inputs passed full hashes; committed source came from Git blobs, excluding every foreign WIP file, including Jedi's WAT/compiler/census and new protected-run-entry test. The task-local polling, breakpoint and relocation adapters were reused and adapted only under ignored scratch/runs/20261103-alice-authentic-ground-gameplay/. Only this note is committed; no production code changed.

Actual native observation

PID 17810, session 9857cf10-1c1a-479e-adc8-5e697dc6d0d6, started 18:13:54.002696Z, with immutable guest/outer deadlines 18:43:54.002696Z / 18:44:54.002696Z, 400-command ceiling, ordinary 100000 instruction budget and 200ms frozen tick. Actual instantiated WASM was 2089591 bytes, SHA256 0a58fd7b9ac82c5303048776f442b15e72528b55ddf85564e0df92fcb1aa384d. The accepted integration build/module was reused after verifying unchanged compile inputs; accepted prerequisite fixtures were not rerun. All original game files, all 27 pngjs package files and actual test/binaries/tlbs/stdole2.tlb were pinned before allocation. Actual startup imports, system-data mounts, fonts, PNG encode/decode and the original fourteen-file companion manifest passed locally and on the boat before original execution.

Personally inspected captures show New Game highlighted, Easy highlighted, letter blocks, the corridor cinematic and Alice upright in the cave with HUD. Only two paired Enter presses selected the menus; no cinematic was skipped. All polls used get-local.py LABEL 0, without resubmission. The longest logo chunk completed after about 124 seconds under its original request deadline.

At 18:24:13.798Z, the earliest native breakpoint was retained in scheduler path cooperative-interleaved: main native thread 1, EIP 00452f63, halt 5, batch 4104, one pending credit, ESP 074f3a0c. The original step remained pending. Status and two raw-span reads retained identical registers, thread census, token and credits. No later scheduler slice admitted guest execution during those reads.

The DLL base was discovered in this session as 02d40000, not assumed from a predecessor. Original engine/DLL full hashes remain 61857b521eb7260c28b3ae6960fa8ee18d4d3679b6190f5221131e382e0f3a72 / 42838c2a96520c4b8b42e710d7d5bcf3a7a7f2a71b141a9f75bff01ee28f3d4c. Offline replay of the authentic captured bytes verifies all 53 original instruction/data windows with individual PE relocations. The native selector also passed its lifecycle/header/signature checks before the contact rejection; its actual lifecycle row values were unfortunately not saved in the failure diagnostic. The repaired adapter now preserves that partial diagnostic context.

The independent retained control identity, breakpoint anchor and captured loopback/qport selector derive slot 0, qport 20818. Engine-client EBX matches the selected client. Record 7ddb0014, state 02bb090c, Player 03aea628, runtime Player vtable, record activity/index, API/engine roots, strides and object/state backlinks agree. Matching state/object copies contain origin (-4183.4814453125, 5999.94970703125, 32.125) and view angles (0, -90, 0) degrees. These are one authentic held observation, not a qualified continuous footing observer or proof of allocation generation.

Concrete remaining failure and bounded repair

capture1 (command37) rejected at 18:24:39.419Z with invalid contact record. The failure preserves 710 read spans / 25385 bytes. Native state has ground number1022, walking1, plane1; Player contact/ground pointers equal records + 1022*2dc. The old decoder incorrectly required 1022 to be less than the live count208 before recognizing reserved world.

Original engine 0045833f..00458363 assigns collision entity1022, uses1023 for the no-hit result, and restores1022 for its solid case. Original DLL 100bcdae..100bcddc excludes indices at/above1022 from the live-count update. The previously established 100e7721..100e7784 contact copy still computes the reserved record pointer. Exact original windows and disassembly are retained in world-contract.json, world-trace-producer.txt and reserved-record-count.txt.

Post-stop contact.cjs recognizes this exact capacity1024 reserved-world case, validates pointer arithmetic and water-dependent copies, and does not dereference it as an active ordinary entity. The authentic old-fail/new-pass contact replay and seven changed-pointer/capacity/flag/index controls pass; three explicitly synthetic absent-contact controls remain distinct. The repaired capture also requires mapped versions of the new original windows and retains partial module/selection diagnostics. That complete repaired capture has not been executed natively. The failed capture ended before reading movement stack arguments; those missing bytes cannot be reconstructed by the offline replay.

A second concrete harness issue appeared during ordinary teardown: the pending step's socket-disconnect error overwrote the legacy first-error.json after the decoder failure had requested stop. Both original per-label receipts and the primary exception remain unchanged. first-failure-analysis.json explains the ordering. Post-stop request-repaired.py uses a serialized first-error writer; eight simultaneous secondary writers plus a later retry preserve the primary bytes and retain every secondary error. This is a passing host-only test, not another native run or a production fault-retention qualification.

Reproduce the focused repairs and authentic observation without a boat:

R=scratch/runs/20261103-alice-authentic-ground-gameplay
node "$R/test-contact.cjs"
node "$R/replay-observation.cjs"
python3 "$R/test-receipt-failure.py"

Syntax checks pass for the adapted CLI, repaired capture and request writer. Post-stop helpers and their full hashes are in post-stop-repair-pins.json. No production code/build input changed or was committed; the preceding accepted locked full build and six focused integration suites are explicitly reused.

Recovery and custody

Original Alice exited 0 by ordinary stop at 18:24:39.519075Z, after 645.516 seconds. All 38 commands are gap-free. No route, Space, mouse, attack, injected guest call, guest-memory write, cheat or config/save edit occurred. The incomplete one-credit request is a teardown disconnect, not a successful sample credit or a native WASM fault. Runtime was released at 18:24:56Z; no further native execution/build followed.

All 6132 remote payloads passed local size/full-SHA verification before the delegated normal STOP. Archive 117710273 bytes, SHA256 2a4a3978ba922febc26d72a0974df74beb7dbba492f362b8519fc4f2485e971f. RECOVERY.md, archive-verification.json, SHA256SUMS, recovered runtime, source pins, raw receipts, screenshots and actual module retain recovery. Own helpers were absent and port8123 closed before STOP; disk stayed above2GB.

Own fresh boat bx_tahsef2b, TTL3600, original expiry 2026-11-03T19:13:29.317Z, was recorded before use and admitted only after Jedi's explicit runtime release. Stop stop_89e248e05034b014 completed 18:26:20.982Z, error null, forced false. Actual state STOPPED at 18:26:20.951Z; available completed snapshot verified 18:26:20.883Z. Custody returns to the coordinator with no active expiry. No foreign or predecessor boat, resume, Claude, HeroesII, CACA003B or public deployment was used. Single task, no subworkers, no successor.

Next: use the retained reserved-world and failure-custody repairs in a bounded original session, complete saved module/stack anchors, establish Player allocation continuity, obtain two neutral supported samples and an ordinary jump/unsupported comparison, then continue the ordinary wind/plank/knife route.